A lot of issues with data augmentation for keypoint datasets

Feedback
1

Hello,

I have a fully annotated dataset but I keep running into various issues when setting it up.

(1) Creating my new dataset results in keypoints appearing that were not annotated in the original image. For example, if I have a keypoint schema with 22 keypoints where only 10 keypoints are visible, the remaining 12 will appear in random positions on the training set images after creating my dataset.

(2) Applying various data augmentations leads to some images having no keypoint annotations. I applied pretty standard data preprocessing and data augmentations to my dataset to triple the size, including: resize (fit), saturation, brightness, and exposure yet when i check the images in my dataset there are some images where keypoints are missing.

I’ve paid quite a bit of money to use the paid version of roboflow and running into these issues has been quire frustrating. Would really appreciate any support here.

2

Hi there,

We’re very sorry you’re experiencing these keypoint dataset issues. We’re looking into a fix.

Have you noticed a specific augmentation that causes the issue, saturation, brightness, or exposure?

Is it ok if we take a look at your workspace versions to diagnose the problem directly?

Best,
Lake

3

Hey, thanks for the response. It does not seem to be related to a specific augmentation. I’ve tried creating datasets with no preprocessing and augmentation and the issue arises.

Yes, feel free to take a look at my workspace.

4

Hey @mz0g, I did some changes and I think it will generate correctly now.

Can you try to generate a new version?

Preprocessing/augs doesn’t matter for this bug.

5

Oh to be clear, I think I fixed (1), but not (2).